AI Briefing

The two loudest doomsayers in AI just quietly admitted the jobs apocalypse hasn't shown up. While Altman and Amodei walk back their predictions ahead of blockbuster IPOs, the actual danger is infrastructure-level: a single-character bug in the Python package that powers most AI agents left millions of servers wide open, and a GitHub tool called Heretic has been stripping safety guardrails from open models faster than labs can ship them. Oh, and four frontier labs acquired four startups in five days, and nobody's calling it a trend yet.

01IndustryFortune ↗

Altman and Amodei Both Walk Back AI Jobs Apocalypse Predictions Ahead of IPOs

OpenAI CEO Sam Altman told a Sydney conference he's 'delighted to be wrong' about AI killing white-collar jobs, admitting he expected far more entry-level job losses by now. Anthropic CEO Dario Amodei, who once predicted AI could eliminate 50% of white-collar work, now says automation may expand the work people do. The reversals come as both companies prepare for massive IPOs, and a Yale Budget Lab study found no meaningful AI-driven rise in unemployment through March 2026. Goldman Sachs CEO David Solomon has been making the same argument since late 2025, and the data is starting to back him up.

02ResearchArs Technica ↗

One-Character Bug in Starlette Exposes Millions of AI Agents to Server Takeover

A critical vulnerability dubbed BadHost (CVE-2026-48710) in Starlette, the open-source framework that underpins FastAPI and receives 325 million downloads per week, allows attackers to bypass path-based authorization by injecting a single character into the HTTP Host header. The bug affects the entire Python AI tooling ecosystem: vLLM, LiteLLM, Text Generation Inference, most OpenAI-shim proxies, MCP servers, agent harnesses, and model-management UIs all depend on Starlette through FastAPI. A fix landed in Starlette 1.0.1, but the sheer breadth of the blast radius means millions of deployments remain vulnerable until they update.

03PolicyThe Irish Times / FT ↗

GitHub Tool 'Heretic' Strips Safety Guardrails from Open AI Models in Under 10 Minutes

An FT investigation found that a freely available GitHub tool called Heretic can remove safety guardrails from Meta's Llama and Google's Gemma models in minutes, no specialist hardware required. The modified models responded to prompts about biological weapons, malware, and child exploitation. Heretic has already been used to create 3,500 'decensored' models with 13 million combined downloads, and its creator stripped safeguards from Google's Gemma 4 within 90 minutes of its release. Google acknowledged that abliteration is a known challenge facing all open models, while Meta declined to comment.

04IndustryTechCrunch ↗

Andrej Karpathy Joins Anthropic to Build AI-Accelerated Pre-training Team

OpenAI co-founder and former Tesla AI chief Andrej Karpathy has joined Anthropic, where he's working on pre-training under team lead Nick Joseph. His mandate is to launch a new team that uses Claude itself to accelerate pretraining research, an increasingly critical frontier as AI companies race to automate parts of AI development. Karpathy is one of the few researchers who bridges LLM theory and large-scale training practice, and his move to Anthropic signals a shift in the AI talent war. The hire comes as Anthropic prepares for its own IPO and has just closed a $30 billion round at a $900B+ valuation.

05IndustryStartupHub.ai ↗

Four Frontier Labs Made Four Acquisitions in Five Days, Signaling Quiet Consolidation

In a single week, Anthropic acquired Stainless (the SDK infrastructure startup that powered developer tooling for OpenAI and Google) for a reported $300M+, Mistral bought Emmi AI (physics-aware industrial AI from Vienna), Google DeepMind hired the entire Contextual AI team via an $80-90M licensing deal, and Meta acqui-hired the Dreamer team. Each deal targeted a different capability gap, and none of the labs framed them as related. But the pattern is clear: frontier labs are buying teams rather than products, structuring deals to avoid antitrust scrutiny, and the startup-to-acquisition pipeline is compressing from years to months.